Liquid AI releases LFM2.5-24B-A2B open-weight model for local hardware

Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model designed to run efficiently on consumer hardware. While the model contains 24 billion total parameters, it utilizes a mixture-of-experts architecture that activates only 2 billion parameters per token, significantly reducing the computational footprint.

The release features an expanded context window of 131,000 tokens and a vocabulary of 128,000 entries to improve support for non-Latin text. The model is optimized for agentic tasks and tool use, incorporating chain-of-thought reasoning. In internal benchmarks, Liquid AI reported high performance in automation, such as scoring 27.69 on AutomationBench, which outperformed competitors like Qwen3.6-35B-A3B.

Designed for local deployment, the model demonstrated high speeds on edge devices, reaching up to 184 tokens per second on Apple’s M5 Max chip. It is available for download on Hugging Face and supports various frameworks including llama.cpp, MLX, and vLLM.